A method and system for predicting short-term load of a transformer district based on space-time correlation

Abstract

The application discloses a transformer area multi-user short-term load prediction method and system based on space-time correlation, and relates to the technical field of power load prediction. The method comprises the following steps: abnormal load data is identified by using a DBSCAN algorithm, and users are clustered by using a K-means algorithm; the mean value of the same time in the cluster is used to replace the abnormal characteristic value, so that an optimized load data set is obtained; a comprehensive correlation analysis index is constructed to quantitatively analyze the correlation between each influencing factor and the load, and the influencing factors with strong correlation are selected as input features of a multi-element load prediction model; characteristics are extracted and predicted in time and space by using an Informer and a GCN respectively, and then a single prediction result is fused by using a variance-covariance combination prediction model to realize load prediction in a future period of time. According to the application, the space-time correlation information of the transformer area multi-user load is introduced, so that the short-term load prediction precision can be effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power load prediction, in particular to a multi-user short-term load prediction method and system based on space-time correlation in a transformer area. BACKGROUND

[0002] Accurate short-term load prediction is essential for the safe and stable operation of the power system and the optimization of control and scheduling. Domestic and foreign scholars have been committed to researching better short-term load prediction methods. Existing researches mainly include traditional time series analysis methods, machine learning-based methods and deep learning-based methods.

[0003] However, with the large-scale access of clean energy to the power grid and the increasing number of electric vehicles, the complexity and uncertainty of the power grid are increasing. In addition, the widespread use of smart meters and other devices at the user side has also brought about the demand for load prediction of a large number of users, which has challenged traditional power load prediction models. Most existing methods model the time series characteristics of a single user and ignore the spatial correlation between multiple users. However, traditional prediction methods that only target a single object cannot meet the new situation of a large number of users. Users in the same area share the same geographical space, weather conditions, electricity price policies and other comprehensive factors. There is often a certain spatio-temporal correlation between user electricity consumption behaviors. If this spatio-temporal correlation can be fully tapped, it will have a very positive significance for the modeling and prediction of short-term user load.

[0004] Therefore, how to tap the spatio-temporal correlation of users and accurately predict the load of a large number of users is a problem that needs to be solved by those skilled in the art. SUMMARY

[0005] Therefore, the present application provides a multi-user short-term load prediction method and system based on space-time correlation in a transformer area, which realizes accurate short-term prediction of multi-user load in a transformer area by introducing the space-time correlation information of multi-user load in the transformer area.

[0006] To achieve the above purpose, the present application provides the following technical solutions:

[0007] The present application discloses a multi-user short-term load prediction method based on space-time correlation in a transformer area, and the specific steps are as follows:

[0008] Obtain a historical load data set of multi-user electricity consumption in a transformer area;

[0009] Identify abnormal load data in the historical load data set by using a DBSCAN algorithm, cluster users by using a K-means algorithm, replace the feature values of abnormal load data with the mean values of the same moment of the class to which the abnormal load data belongs, and obtain an optimized load data set;

[0010] The correlation between the influence factors and the load in the optimized load data set is quantitatively analyzed by using a comprehensive correlation analysis index, and the influence factors with strong correlation are selected to obtain a predicted load data set as an input of the multivariate load prediction model.

[0011] The multivariate load prediction model uses an Informer model and a GCN model for feature extraction and prediction, and then fuses the prediction results of the Informer model and the GCN model through a variance-covariance combination prediction model to obtain a final load prediction result.

[0012] Further, the specific steps for obtaining the optimized load data set are as follows:

[0013] Step 1: In the historical load data set, the measured load data of all users in a t time period in a substation is intercepted to construct a sample matrix, which is represented as:

[0014]

[0015] wherein, P i,j represents the measured load of user i at the jth moment, P i represents a sample group of the sample matrix P, m represents the total number of users in the substation, and k represents the number of sampling points.

[0016] Step 2: All sample groups are divided into c user groups Group a (a = 1, 2, …, c) by using a K-means algorithm.

[0017] Step 3: The mean load of each user group at the same moment is used to replace the eigenvalue of the abnormal load data.

[0018] Further, the step 2 specifically includes:

[0019] Step 2.1: c initial cluster centers P a are randomly selected from all sample groups.

[0020] Step 2.2: The Euclidean distance d of each sample group to each initial cluster center is calculated, and the formula is:

[0021]

[0022] Step 2.3: Any sample group P i is divided into the class represented by the nearest cluster center P a , and all sample groups are traversed to complete one division.

[0023] Step 2.4: The sample mean of each class is calculated, and the total sum of squared errors of clustering is calculated:

[0024]

[0025] in, N represents the sample mean of the a-th user group; a Let be the total number of samples in the a-th class; E is the sum of the squared clustering errors;

[0026] Step 2.5: Using the mean of the samples in each class as the new cluster center, return to step 2.2 for iterative clustering until the sum of the squared clustering errors converges, and the final clustering result is obtained.

[0027] Furthermore, the comprehensive correlation analysis index C M The formula is:

[0028]

[0029] Where, ρ X,Y M is the Pearson correlation coefficient value. IC This represents the maximum information coefficient value.

[0030] When C M When the value is greater than the preset threshold σ, it indicates that there is a strong correlation between the influencing factor variable X and the load Y. Influencing factors that are strongly correlated with load forecasting are selected as the input feature variables of the final load combination forecasting model.

[0031] Furthermore, the Pearson correlation coefficient value ρ X,Y The calculation formula is expressed as:

[0032]

[0033] Among them, Y i X is the actual observed data sequence of the load. i Let l be the sequence of characteristic data that may affect load forecasting, and ρ be the length of the sequence. X,Y This represents the Pearson correlation coefficient between variable X and load Y.

[0034] The maximum information coefficient value M IC The calculation formula is expressed as:

[0035]

[0036] Wherein I(X) i ,Y j ) is X i With Y i Mutual information between them, |X i | and |Y i |representing sequence X i With Y i The length, min(|Xi |,|Y i |) indicates taking sequence X i With Y i The minimum length.

[0037] Furthermore, the Informer model selects load data from users at H time steps from the predicted load dataset as input, predicts the load at time p in the future, and obtains the prediction output.

[0038] Furthermore, the GCN model uses a weighted user adjacency matrix A and a user feature matrix B as input to predict the load at time p in the future, and obtains the output result.

[0039] Furthermore, the weighted user adjacency matrix A is represented as:

[0040] A={R i1 ,R i2 ,…,R if};

[0041] Among them, R if The spatial correlation between user i and user f is calculated using the Pearson correlation coefficient method, and the formula is as follows:

[0042]

[0043] Among them, P i The actual observed data sequence for user i's load; P f Let l be the actual observed data sequence of user f's load, where l is the sequence length;

[0044] The user feature matrix B is represented as follows:

[0045]

[0046] Among them, P i,j This represents the measured load value of user i at a certain historical moment j.

[0047] Furthermore, the variance-covariance combined prediction model calculates the combined weight coefficients of the Informer model and the GCN model respectively, and outputs the final load forecast values ​​for the next p time periods based on the calculated combined weights; the variance-covariance combined prediction model is expressed as:

[0048]

[0049] in y represents the sequence of load forecast values ​​for the next p times predicted by the Informer model and the GCN model, respectively. t w represents the sequence of the final load forecast values. T w S They are respectively The weighting coefficients, and w T +w S =1;

[0050] Weighting coefficient w T w S The method for determining it is as follows:

[0051] Based on the predicted load dataset, the corresponding prediction results are calculated using the Informer model, the GCN model, and the variance-covariance combined prediction model. and y t ;

[0052] A sequence is formed by selecting measured load data at time t from the predicted load dataset. Calculate the error sequences for the three prediction results respectively:

[0053]

[0054] in, e t They are respectively y t The corresponding error sequence;

[0055] The relationship between the weighting coefficients and the error sequence is expressed as follows:

[0056]

[0057] The error variance of the final load forecast sequence is expressed as:

[0058]

[0059] Among them, Var(e t ) represents the error variance of the final load forecast sequence. for The covariance of w; T For Var(e) t Finding the minimum value, we get:

[0060]

[0061] w S =1-w T .

[0062] The application also discloses a transformer area multi-user short-term load prediction system based on space-time correlation.

[0063] A data acquisition module is configured to acquire transformer area multi-user historical load data sets.

[0064] A data processing module is configured to identify abnormal load data in the historical load data sets by using a DBSCAN algorithm, cluster users by using a K-means algorithm, replace feature values of the abnormal load data with in-class mean values of the same moment of the class to which the abnormal load data belongs, and obtain an optimized load data set.

[0065] A feature optimization module is configured to quantitatively analyze the correlation between each load influencing factor and load in the optimized load data set by using a comprehensive correlation analysis index, select the influencing factors with strong correlation to obtain a predicted load data set as input of a multi-element load prediction model.

[0066] A load prediction module is configured to use an Informer model and a GCN model to extract features and make predictions, and fuse the prediction results of the Informer model and the GCN model by using a variance-covariance combination prediction model to obtain a final load prediction result.

[0067] Compared with the prior art, the application discloses a transformer area multi-user short-term load prediction method and system based on space-time correlation, which can effectively improve the short-term load prediction accuracy by introducing the space-time correlation information of the transformer area multi-user load. The abnormal data correction method based on K-means spatial clustering uses the measured load as a feature to cluster users, and replaces the abnormal feature values with the in-class mean values of the same moment, which can effectively fuse the space-time features to improve the abnormal data correction accuracy. The comprehensive correlation analysis index is used to quantitatively analyze the correlation between each influencing factor and load, and the features with strong correlation are selected as the input of the model, which can simplify the model and data while effectively improving the prediction accuracy and efficiency. The Informer is used to extract time features, the GCN is used to extract space features, and then the space-time features are fused by using the variance-covariance combination prediction method to complete the final load prediction, which can fully utilize the space-time information to improve the load prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0068] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only embodiments of the application, and those skilled in the art can obtain other drawings according to the provided drawings without creative labor.

[0069] Figure 1 The figure is a whole flowchart of the embodiment of the present application.

[0070] Figure 2 The figure is a structure diagram of the Informer model of the embodiment of the present application.

[0071] Figure 3 The figure is a structure diagram of the GCN model of the embodiment of the present application.

[0072] Figure 4 The figure is a load combination prediction flowchart based on Informer and GCN of the embodiment of the present application. DETAILED DESCRIPTION

[0073] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0074] The embodiment of the present application discloses a transformer area multi-user short-term load prediction method based on spatio-temporal correlation, as shown in the figure, and the specific steps are as follows: Figure 1

[0075] Obtain a transformer area multi-user electricity historical load data set;

[0076] Use the DBSCAN algorithm to identify the abnormal load data in the historical load data set, cluster the users through the K-means algorithm, replace the feature values of the abnormal load data with the in-class mean values of the same moment of the class to which the abnormal load data belongs, and obtain an optimized load data set;

[0077] Use the comprehensive correlation analysis index to quantitatively analyze the correlation between each load influencing factor and load in the optimized load data set, select the strong correlation influencing factor to obtain a predicted load data set, and use it as the input of the multi-element load prediction model;

[0078] The multi-element load prediction model uses the Informer model and the GCN model for feature extraction and prediction, and then uses the variance-covariance combination prediction model to fuse the prediction results of the Informer model and the GCN model to obtain the final load prediction result.

[0079] In a specific embodiment, the specific steps for obtaining the optimized load data set are as follows:

[0080] ​Step 1, in the historical load data set, intercept all users in the past t time period within the measured load data to construct a sample matrix, the sample matrix is represented as:

[0081]

[0082] Where, P i,j represents the measured load of user i at the jth moment, P i represents a sample group of sample matrix P, m represents the total number of users in the area, and k represents the number of sampling points.

[0083] Step 2, all sample groups are divided into c user groups Group a (a = 1, 2, …, c) by K-means algorithm.

[0084] Step 3, the mean load of each user group at the same moment is used to replace the eigenvalue of abnormal load data.

[0085] In a specific embodiment, step 2 specifically includes:

[0086] Step 2.1, randomly select c initial cluster centers P a from all sample groups.

[0087] Step 2.2, calculate the Euclidean distance d of each sample group to each initial cluster center, the formula is:

[0088]

[0089] Step 2.3, any sample group P i is divided into the class represented by the nearest cluster center P a , and all sample groups are traversed to complete a division;

[0090] Step 2.4, calculate the sample mean of each class, and calculate the total sum of squared errors of clustering:

[0091]

[0092] Where, is the sample mean of the a th user group; N a is the total number of samples in the a th class; E is the total sum of squared errors of clustering.

[0093] Step 2.5, take the sample mean of each class as the new cluster center, return to step 2.2 for iterative clustering division until the total sum of squared errors of clustering converges, and the final clustering result is obtained.

[0094] In one specific embodiment, DBSCAN algorithm is used to identify abnormal data. DBSCAN algorithm is a density-based noise space clustering algorithm, which reflects the tightness of data distribution through two core parameters, neighborhood radius ε and minimum sample point number MinPts, and can cluster densely packed scattered points (i.e. high-density data region) and regard discrete points as noise points, which is very suitable for abnormal load data identification. The specific steps are as follows:

[0095] For the obtained historical load data set, set the neighborhood radius ε and the minimum sample point number MinPts, ε(p) represents the neighborhood with p as the center and ε as the radius, and MinPts represents the number of sample points contained in ε(p); randomly select a sample point p from the historical load data set, if the neighborhood ε(p) of p contains at least MinPts data points, mark the point as a core object; if the neighborhood ε(p) of p contains less than MinPts data points and p is not located in other neighborhoods, mark the point as a noise point, i.e. abnormal data; randomly select an unprocessed point, repeat the noise point judgment operation until all sample points in the historical load data set are processed.

[0096] Generally, the neighborhood radius ε and the minimum sample point number MinPts are determined by k-dist graph method, and the specific steps are as follows:

[0097] Calculate the Euclidean distance between each object p i and the k-th nearest object p e of the object, denoted as the k-dist value of the object, i.e.

[0098]

[0099] Where n represents the data dimension, i.e. the number of features (including meteorological factor data, time characteristics, historical load data, etc., such as temperature, air pressure, precipitation, holidays, etc.);

[0100] Arrange all object k-dist values in ascending or descending order and draw k-dist curve;

[0101] Take the distance corresponding to the first "inflection point" of the k-dist curve as the neighborhood radius ε, and take k+1 as the minimum sample point number MinPts (to reduce the amount of calculation, generally k=2n-1 is used by default).

[0102] In one specific embodiment, step 3 replaces the abnormal load data with the mean value of the same time class, and the specific formula is as follows:

[0103]

[0104] Wherein, represents an abnormal data correction value; N a represents the total number of samples of the a-th class, represents the measured value of the n-th feature of the user f at the j-th time point.

[0105] In one specific embodiment, the comprehensive correlation analysis index C M is calculated according to the following formula:

[0106]

[0107] wherein, ρ X,Y is a Pearson correlation coefficient value, and M IC is a maximum information coefficient value.

[0108] When C M is greater than a preset threshold σ, it indicates that there is a strong correlation between the influence factor variable X and the load Y, and the influence factor having a strong correlation with the load prediction is selected as an input feature variable of the final load combination prediction model.

[0109] In one specific embodiment, the calculation formula of the Pearson correlation coefficient value ρ X,Y is as follows:

[0110]

[0111] wherein, Y i is an actual observation data sequence of the load, X i is a feature data sequence that may affect the load prediction, l is the length of the sequence, and ρ X,Y represents the Pearson correlation coefficient value between the variable X and the load Y.

[0112] The calculation formula of the maximum information coefficient value M IC is as follows:

[0113]

[0114] wherein, I(X i , Y j ) is the mutual information between X i and Y i , |X i | and |Y i | are the lengths of the sequences X i and Y i , and min(|X i |, |Y i |) represents the minimum value of the lengths of the sequences X i and Y i .

[0115] The embodiment of the application utilizes Informer and GCN to respectively perform feature extraction and prediction in time and space, and then fuses single prediction results through a variance-covariance combination prediction model to realize load prediction in a certain period of time in the future. Taking the Informer-GCN combination prediction model as a multivariate load prediction model, the model mainly includes three modules: a time feature extraction and prediction module, a space feature extraction and prediction module, and a time-space feature fusion and prediction module, and the model structure is as shown in Figure 4 .

[0116] In a specific embodiment, the time feature extraction and prediction module is mainly composed of Informer. Informer is a deep network model suitable for long time series prediction. It builds a deep network through an attention mechanism to extract long time series long-time dependency relationships, and the structure is as shown in Figure 2 . The entire model is divided into an encoder and a decoder. The encoder uses a position embedding layer to perform time coding on the input sequence, and then uses multiple neural network modules composed of multi-head probabilistic sparse self-attention and convolutional layers for deep feature extraction. The decoder inputs the load-related variable information, which is operated with the deep feature map extracted by the encoder to capture the correlation between the features, and finally obtains the estimated load prediction value through a fully connected layer. In this module, the Informer model selects H time steps of load data of users in the load data set as input (including multiple variable sequences such as unit state data, weather data, etc.), predicts the load at time p in the future, and obtains the prediction output result

[0117] In a specific embodiment, the space feature extraction and prediction module is mainly composed of GCN. GCN is a deep learning method for processing non-Euclidean space graph data. It aggregates the features of itself and neighboring nodes in the adjacency matrix through convolution operation to generate a new node feature vector, and the structure is as shown in Figure 3 . In this module, the GCN model takes the weighted user adjacency matrix A and the user feature matrix B as input to predict the load at time p in the future, and obtains the output result

[0118] In a specific embodiment, the weighted user adjacency matrix A is represented as:

[0119] A={R i1 ,R i2 ,…,R if};

[0120] Wherein, R if represents the spatial correlation degree of user i and user f, which is calculated according to the Pearson correlation coefficient method, and the calculation formula is represented as:

[0121]

[0122] Among them, P i The actual observed data sequence for user i's load; P f Let l be the actual observed data sequence of user f's load, where l is the sequence length;

[0123] The user feature matrix B is represented as follows:

[0124]

[0125] Among them, P i,j This represents the load measured by user i at time j.

[0126] In a specific embodiment, the variance-covariance combined forecasting model calculates the combined weight coefficients of the Informer model and the GCN model respectively, and outputs the final load forecast values ​​at p future time points based on the calculated combined weights; the variance-covariance combined forecasting model is expressed as:

[0127]

[0128] in y represents the sequence of load forecasts for the next p times, predicted by the Informer model and the GCN model, respectively. t w represents the sequence of final load forecasts T w S They are respectively The weighting coefficients, and w T +w S =1;

[0129] Weighting coefficient w T w S The method for determining it is as follows:

[0130] Based on the predicted load dataset, the prediction results were calculated using the Informer model, the GCN model, and the variance-covariance combined prediction model. and y t ;

[0131] Select measured load data at time t from the predicted load dataset to form a sequence. Calculate the error sequences for the three prediction results respectively:

[0132]

[0133] in, e t They are respectively y t The corresponding error sequence;

[0134] The relationship between the weight coefficient and the error sequence is expressed as:

[0135]

[0136] The error variance of the final load prediction sequence is expressed as:

[0137]

[0138] Wherein, Var(e t ) is the error variance of the final load prediction sequence, is the covariance of ; about w T Minimizing Var(e t ) obtains:

[0139]

[0140] w S =1-w T .

[0141] The embodiment of the application also discloses a transformer area multi-user short-term load prediction system based on space-time correlation, comprising:

[0142] A data acquisition module is configured to acquire transformer area multi-user historical load data set;

[0143] A data processing module is configured to identify abnormal load data in the historical load data set by using a DBSCAN algorithm, cluster users by using a K-means algorithm, replace feature values of the abnormal load data with in-class mean values of the same moment of the class to which the abnormal load data belongs, and obtain an optimized load data set;

[0144] A feature optimization module is configured to use a comprehensive correlation analysis index to quantitatively analyze the relevance between each load influencing factor and load in the optimized load data set, select the influencing factor with strong correlation to obtain a predicted load data set as an input of a multi-element load prediction model;

[0145] A load prediction module is configured to use an Informer model and a GCN model for feature extraction and prediction, and fuse prediction results of the Informer model and the GCN model by using a variance-covariance combination prediction model to obtain a final load prediction result.

[0146] The various embodiments described in this specification are implemented in a progressive manner, each embodiment focusing on the differences from other embodiments, and the same or similar parts between embodiments can be mutually referred to. For the apparatus disclosed by the embodiments, since it corresponds to the method disclosed by the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0147] The above description of disclosed embodiments enables one of ordinary skill in the art to make or use the application. Various modifications to these embodiments will be readily apparent to those of ordinary skill in the art, and the generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the application. Therefore, the application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for short-term load forecasting of multiple users in a distribution area based on spatiotemporal correlation, characterized in that, The specific steps are as follows: Obtain historical load data of electricity consumption for multiple users in the transformer area; The DBSCAN algorithm is used to identify abnormal load data in the historical load dataset. The K-means algorithm is used to cluster users. The feature values ​​of the abnormal load data are replaced by the intra-cluster mean of the class to which the abnormal load data belongs, so as to obtain the optimized load dataset. Using comprehensive correlation analysis indicators, the correlation between each load influencing factor and the load in the optimized load dataset is quantitatively analyzed, and the influencing factors with strong correlation are selected to obtain the predicted load dataset, which is used as the input of the multivariate load prediction model. The multivariate load forecasting model uses the Informer model and the GCN model for feature extraction and forecasting, and then uses the variance-covariance combined forecasting model to fuse the forecasting results of the Informer model and the GCN model to obtain the final load forecasting result. The Informer model selects load data from users at H time steps from the predicted load dataset as input, and then... p The load at a given time is predicted to obtain the predicted output. ; The GCN model uses a weighted user adjacency matrix. and user feature matrix As input, for the future p The load at a given time is predicted to obtain the output result. ; The weighted user adjacency matrix Represented as: ； in, On behalf of users With users The spatial correlation degree is calculated using the Pearson correlation coefficient method, and the calculation formula is as follows: ； in, For users Actual observation data sequence of load; For users Actual observed data sequence of load, The sequence length; The user feature matrix B Represented as: ； in, On behalf of users A certain history The measured load value at any given time; The variance-covariance combined prediction model calculates the combined weight coefficients of the Informer model and the GCN model respectively, and outputs the future based on the calculated combined weights. The final load forecast value at each time point; the variance-covariance combined forecast model is expressed as: ； in , These respectively represent the Informer model and the GCN model predicting the future. A sequence of load forecast values ​​at each time point. This represents a sequence consisting of the final load forecast values. , They are respectively , The weighting coefficients, and ; Weighting coefficient , The method for determining it is as follows: Based on the predicted load dataset, the corresponding prediction results are calculated using the Informer model, the GCN model, and the variance-covariance combined prediction model. , and ; Select from the predicted load dataset The sequence of measured load data at different times Calculate the error sequences for the three prediction results respectively: ； ； ； in, , , They are respectively , , The corresponding error sequence; The relationship between the weighting coefficients and the error sequence is expressed as follows: ； The error variance of the final load forecast sequence is expressed as: ； in, The error variance of the final load forecast sequence. for Covariance; regarding right Finding the minimum value, we get: ； 。 2. The method for short-term load forecasting of multiple users in a distribution area based on spatiotemporal correlation according to claim 1, characterized in that, The specific steps for obtaining the optimized load dataset are as follows: Step 1: From the historical load dataset, extract the measured load data of all users within the transformer area over the past time period t to construct a sample matrix. The sample matrix is ​​represented as follows: ； in, Indicates user In the The load measured at a given time. Represents the sample matrix A sample group, This indicates the total number of users within the region. Indicates the number of sampling points; Step 2: Divide all sample groups into c user groups using the K-means algorithm. ; Step 3: Replace the feature value of the abnormal load data with the average load value of each user group at the same time.

3. The method for short-term load forecasting of multiple users in a distribution area based on spatiotemporal correlation according to claim 2, characterized in that, Step 2 specifically includes: Step 2.1: Randomly select c initial cluster centers from all sample groups. ; Step 2.2: Calculate the Euclidean distance from each sample group to each initial cluster center. d The formula is: ； Step 2.3, select any sample group Assign to the nearest cluster center Within the class it represents, iterate through all sample groups to complete one partition; Step 2.4: Calculate the sample mean for each class and the sum of squared errors for clustering. ； ； in, For the first a The sample mean of each user group; For the first a The total number of samples in each class; This represents the sum of the squared clustering errors; Step 2.5: Using the mean of the samples in each class as the new cluster center, return to step 2.2 for iterative clustering until the sum of the squared clustering errors converges, and the final clustering result is obtained.

4. The method for short-term load forecasting of multiple users in a distribution area based on spatiotemporal correlation according to claim 1, characterized in that, The comprehensive correlation analysis indicators The formula is: ； in, The Pearson correlation coefficient value. This represents the maximum information coefficient value. when Greater than the preset threshold When, it represents the influencing factor variable. With load The factors that are strongly correlated with load forecasting are selected as the input feature variables of the final load combination forecasting model.

5. A method for short-term load forecasting of multiple users in a distribution area based on spatiotemporal correlation according to claim 4, characterized in that, The Pearson correlation coefficient value The calculation formula is expressed as: ； in, This is the actual observed data sequence of the load. For various characteristic data sequences that may affect load forecasting, The length of the sequence. Representing variables With load The Pearson correlation coefficient between them; The maximum information coefficient value The calculation formula is expressed as: ； in, for and Mutual information between them and Sequences and Length, Indicates taking the sequence and The minimum length.

6. A short-term load forecasting system for multiple users in a distribution area based on spatiotemporal correlation, employing the short-term load forecasting method for multiple users in a distribution area based on spatiotemporal correlation as described in any one of claims 1-5, characterized in that, include: Data acquisition module: Acquires historical load data of electricity consumption for multiple users in the transformer area; Data processing module: The DBSCAN algorithm is used to identify abnormal load data in the historical load dataset. The K-means algorithm is used to cluster users. The feature values ​​of the abnormal load data are replaced with the intra-cluster mean of the class to which the abnormal load data belongs, so as to obtain the optimized load dataset. Feature optimization module: Using comprehensive correlation analysis indicators, the correlation between each load influencing factor and the load in the optimized load dataset is quantitatively analyzed, and the influencing factors with strong correlation are selected to obtain the predicted load dataset, which is used as the input of the multivariate load prediction model; Load forecasting module: The multivariate load forecasting model uses the Informer model and the GCN model for feature extraction and forecasting, and then uses the variance-covariance combined forecasting model to fuse the forecasting results of the Informer model and the GCN model to obtain the final load forecasting result.

Images